diff --git a/megatron/training/checkpointing.py b/megatron/training/checkpointing.py
index 93e8a8acd6a..b73466ccfde 100644
--- a/megatron/training/checkpointing.py
+++ b/megatron/training/checkpointing.py
@@ -1449,10 +1449,10 @@ def _set_arg(arg_name, old_arg_name=None, force=False):
     _set_arg('moe_latent_size', force=True)
 
     # Tokenizer args.
-    _set_arg('tokenizer_type', force=True)
     # Using checkpoint version might not always be safe (e.g., if running on different cluster).
     if args.use_tokenizer_model_from_checkpoint_args:
         _set_arg('tokenizer_model', force=True)
+        _set_arg('tokenizer_type', force=True)
     _set_arg('tiktoken_pattern', force=True)
     _set_arg('padded_vocab_size')